What: The news anchor is GLM-5.2, Zhipu AI's open-weights model that just topped the Artificial Analysis Intelligence Index; the concept it makes concrete is active vs total parameters — the two numbers in its "744B total / 40B active" spec.
Why: Those two numbers price two different things: total sets the memory footprint and the GPU you need, while active sets the compute and bandwidth you pay per token. Reading both tells you what a model release actually costs to run.
vs prior: The old habit was to quote one parameter count — which assumes a dense model where every weight fires on every token, so active equals total. A sparse Mixture-of-Experts splits that into two, and the gap between them is the design lever.
Think of it as
A big engine that fires only a few of its cylinders at a time.
